Dense Pedestrian Detection Methods Based On Improved YOLOv5

Pedestrian detection is the focus of object detection in computer vision.The current pedestrian detection methods based on deep learning have been able to achieve good recognition and detection in simple scenes.However,due to occlusion and variable pedestrian scales in dense scenes,the dense pedestrian detection model cannot fully extract and aggregate pedestrian features,resulting in false detection and missing detection.Therefore,this paper investigates the above problems based on YOLOv5,and proves that the proposed method in this paper can solve the above problems well through experiments.The details are as follows:1.For the problems of occlusion in dense scenes and the weak feature extraction ability of YOLOv5,three attention mechanism fusion schemes are proposed.The attention mechanism is connected with the channel attention submodule and the spatial attention submodule respectively along two independent dimensions,and then the attention information of the two dimensions is summarized to obtain more reliable dense pedestrian attention information.The MR（Miss Rite,MR）of the best fusion scheme CBAM＿C on the Wider Person dataset is reduced to 37.40,and the AP_0.5and AP_0.5:0.95are increased by 0.37 and 1.38,respectively,and also achieved better detection results on the City Person dataset.2.To address the problems of variable pedestrian scale in dense scenes and the weak feature aggregation ability of YOLOv5,based on the above research,the Focus Multihead-YOLO dense pedestrian detection method is proposed based on the BIFPN（Bidirectional Feature Pyramid Network,BIFPN）.The specific improvement are as follows:Remove the single input node in the original feature aggregation network to simplify the network structure without degrading the model performance;Establish spanning connections between the input and output nodes in the feature aggregation network,and fuse more dense pedestrian features without increasing the excessive cost;Set a learnable weight coefficient on each input node in the feature aggregation network,and the size of the coefficient represents the contribution of the input node to the network,so that the network can set the corresponding attention according to the size of the coefficient;Increase the number of detection layers,add a 10*10 size detection layer to the original three detection layers,in order to enrich the dense pedestrian detection scale and reduce the rate of missed detection.The experimental results show that the MR of this method is reduced to 36.34 on the Wider Person dataset,improved by 0.92 and 2.09 on AP_0.5and AP_0.5:0.95,respectively,and reduced by2.72 on the City Person dataset,improved by 4.54 and 1.29 on AP_0.5and AP_0.5:0.95,respectively.3.In order to obtain more dense pedestrian characteristics and find a suitable bounding box regression loss function for dense pedestrian detection tasks,based on Focus Multihead-YOLO,this paper presents a comparative analysis of three bounding box regression loss functions for Io U（Intersection over Union,Io U）and its deformation.Based on Alpha-intersection over Union bounding box regression loss function,the model loss function was improved,and the Focus Multihead Adaptive-YOLO dense pedestrian detection method was proposed.The experiments show that the MR is reduced by 2.42 on the Wider Person dataset,and improved by 2.02 and 3.47 on AP_0.5and AP_0.5:0.95respectively,and also has significant effect on the City Person dataset.Figure[30]Table[11]Reference[55]...